Function retrieval method for intelligent agent office system

By converting user search instructions into readable form in an intelligent office system, and using large language models to automatically search the required content, the problems of low efficiency and insufficient accuracy of function search in the office system are solved, and convenient and efficient function search is achieved .

CN119025550APending Publication Date: 2024-11-26BEIYIN FINANCIAL TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410905234.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-08
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

The function retrieval methods of existing office systems are inefficient and insufficiently accurate. Users need to change keywords multiple times or write complex SQL query statements to retrieve the required functions, resulting in high labor costs and time costs.

Method used

By obtaining the user's first search instruction, using preset instructions to convert a model (such as a pre-trained language model or a neural network language model) to convert it into a second search instruction readable by the agent's office system, and automatically drives the system to retrieve the required content. Reduce user operation complexity.

Benefits of technology

It realizes efficient and accurate function retrieval in the intelligent office system, reduces manual operation costs and time, and improves search convenience and accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119025550A_ABST
    Figure CN119025550A_ABST
Patent Text Reader

Abstract

The invention provides a function retrieval method for an intelligent agent office system, which comprises the following steps: acquiring a first retrieval instruction for the intelligent agent office system, the first retrieval instruction being used for reflecting the demand content of a user for the intelligent agent office system; the first retrieval instruction is converted into input information of a preset instruction conversion model, and the preset instruction conversion model is used for reflecting the incidence relation between the demand content and a second retrieval instruction for obtaining the demand content in the intelligent agent office system; determining a second retrieval instruction based on the input information and a preset instruction conversion model; and inputting the second retrieval instruction into the intelligent agent office system, so that the intelligent agent office system provides the demand content based on the second retrieval instruction. According to the technical scheme, convenience and accuracy of internal retrieval of the office system can be considered, retrieval efficiency can be improved, and retrieval difficulty and cost are synchronously reduced.
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present application relates to the field of information retrieval technology, and in particular to a function retrieval method for an intelligent office system. [Background technology]

[0002] With the rapid development of science and technology, enterprises have increasing demands for information management and office automation. In traditional office systems, when users need to search for a certain function, they often rely on manual operation to enter keywords in the office system for retrieval. However, the keywords provided by the user may be different from the office system's own definition of the functions required by the user. Users often need to change keywords multiple times to retrieve the required functions. Therefore, the keyword search method is difficult to efficiently and accurately meet complex office needs.

[0003] In this regard, the related art proposes to establish an enterprise information database association query system, store data in a structured form in a relational database, and allow users to retrieve information by writing SQL query statements. However, writing SQL query statements is not simple, and adds preparatory work for user retrieval.

[0004] In addition, a rule-based automated office system has been proposed in the related art, where business personnel design a series of rules and logical conditions for information management and task automation based on business needs and data characteristics. On this basis, the system retrieves the specific instructions after the user re-enters the specific instructions in the system. However, this solution is based on the user's understanding of the specific instructions for each function in the system. Even if the specific instructions for each function in the system are listed, the user still needs to find the required instructions in the listed specific instructions. Therefore, this method still requires high labor costs and time costs, and it is difficult to efficiently meet complex office needs.

[0005] Therefore, how to efficiently and accurately retrieve the functions required by users in the office system has become a technical problem that needs to be solved urgently. [Summary of the invention]

[0006] The embodiment of the present application provides a function retrieval method for an intelligent office system, aiming to solve the technical problem that the function retrieval method provided by the office system in the related art is inefficient and lacks accuracy.

[0007] In a first aspect, an embodiment of the present application provides a function retrieval method for an intelligent office system, comprising:

[0008] Acquire a first search instruction for the intelligent office system, wherein the first search instruction is used to reflect the user's demand content for the intelligent office system, and the demand content includes executing the operation action requested by the first search instruction and / or outputting the output content requested by the first search instruction;

[0009] Converting the first search instruction into input information of a preset instruction conversion model, wherein the preset instruction conversion model is used to reflect the association relationship between the demand content and a second search instruction for obtaining the demand content in the agent office system;

[0010] Determining the second search instruction based on the input information and the preset instruction conversion model;

[0011] The second search instruction is input into the intelligent agent office system, so that the intelligent agent office system provides the required content based on the second search instruction.

[0012] In one embodiment of the present application, optionally, converting the first search instruction into input information of a preset instruction conversion model includes:

[0013] Entities in the first search instruction are extracted as input information of the preset instruction conversion model.

[0014] In one embodiment of the present application, optionally, converting the first search instruction into input information of a preset instruction conversion model includes:

[0015] Preset prefix information is added to the first search instruction to obtain input information of a preset instruction conversion model, wherein the preset prefix information is used to describe the processing requirements for converting the first search instruction into a second search instruction readable by the intelligent office system.

[0016] In one embodiment of the present application, optionally, the preset instruction conversion model is a pre-trained language model, a neural network language model, a deep learning language model or a hybrid language model.

[0017] In one embodiment of the present application, optionally, before determining the second search instruction based on the input information and the preset instruction conversion model, the method further includes:

[0018] Performing data structuring processing on the input information to obtain updated input information matching a preset structured format;

[0019] The determining the second search instruction based on the input information and the preset instruction conversion model includes:

[0020] The second search instruction is determined based on the updated input information and the preset instruction conversion model.

[0021] In one embodiment of the present application, optionally, before determining the second search instruction based on the input information and the preset instruction conversion model, the method further includes:

[0022] Based on the initial sample data set, train the initial instruction conversion model;

[0023] Get a dynamically updated real-time sample dataset;

[0024] In response to each update operation of the real-time sample data set, based on the updated real-time sample data set, adjusting the top-level parameters of the initial instruction conversion model while freezing the bottom-level parameters of the initial instruction conversion model to obtain an iterative instruction conversion model;

[0025] After the number of adjustments reaches the preset upper limit, the last iteration instruction conversion model is set as the preset instruction conversion model.

[0026] In one embodiment of the present application, optionally, the method further includes:

[0027] After the agent office system provides the required content based on the second search instruction in multiple operations, determining the accuracy, recall, precision and F1 score of the multiple operations;

[0028] Normalizing the accuracy, the recall, the precision, and the F1 score;

[0029] Based on the result of the normalization processing, determining a set of correction parameters, wherein the set of correction parameters includes a weighted average of normalized values ​​of each of the multiple information in the accuracy, the recall, the precision and the F1 score;

[0030] Based on the historical model confidence and the set of correction parameters, a current confidence of the preset instruction conversion model is determined.

[0031] In one embodiment of the present application, optionally, determining the current confidence of the preset instruction conversion model based on the historical model confidence and the correction parameter set includes:

[0032] Obtaining the variance of the correction parameter set;

[0033] The difference between the variance and the historical model confidence is set as the current confidence of the preset instruction conversion model.

[0034] In a second aspect, an embodiment of the present application provides an intelligent operation device for an intelligent office system, including:

[0035] A first search instruction acquisition unit is used to acquire a first search instruction for the intelligent office system, wherein the first search instruction is used to reflect the user's demand content for the intelligent office system, and the demand content includes executing the operation action requested by the first search instruction and / or outputting the output content requested by the first search instruction;

[0036] An input information determination unit, configured to convert the first search instruction into input information of a preset instruction conversion model, wherein the preset instruction conversion model is used to reflect the association relationship between the demand content and a second search instruction for obtaining the demand content in the agent office system;

[0037] A second search instruction acquisition unit, configured to determine the second search instruction based on the input information and the preset instruction conversion model;

[0038] The required content output unit is used to input the second search instruction into the intelligent agent office system, so that the intelligent agent office system provides the required content based on the second search instruction.

[0039] In one embodiment of the present application, optionally, the input information determining unit includes:

[0040] The first execution unit is used to extract entities in the first search instruction as input information of the preset instruction conversion model.

[0041] In one embodiment of the present application, optionally, the input information determining unit includes:

[0042] The second execution unit is used to add preset prefix information to the first search instruction to obtain input information of the preset instruction conversion model, wherein the preset prefix information is used to describe the processing requirements for converting the first search instruction into a second search instruction readable by the intelligent office system.

[0043] In one embodiment of the present application, optionally, the preset instruction conversion model is a pre-trained language model, a neural network language model, a deep learning language model or a hybrid language model.

[0044] In one embodiment of the present application, optionally, the device further includes:

[0045] An input information structured processing unit, used for performing data structured processing on the input information before determining the second search instruction, to obtain updated input information matching a preset structured format;

[0046] The second search instruction acquisition unit is used for:

[0047] The second search instruction is determined based on the updated input information and the preset instruction conversion model.

[0048] In one embodiment of the present application, optionally, the device further includes:

[0049] An initial model training unit, used for training an initial instruction conversion model based on an initial sample data set before determining the second retrieval instruction;

[0050] A real-time sample data set acquisition unit, used to acquire a dynamically updated real-time sample data set;

[0051] a model parameter optimization unit, configured to adjust the top-level parameters of the initial instruction conversion model in response to each update operation of the real-time sample data set and based on the updated real-time sample data set, while freezing the bottom-level parameters of the initial instruction conversion model, to obtain an iterative instruction conversion model;

[0052] The model determination unit is used to set the last iteration instruction conversion model as the preset instruction conversion model after the adjustment times reach the preset upper limit.

[0053] In one embodiment of the present application, optionally, the device further includes:

[0054] A feedback information determination unit, configured to determine the accuracy, recall, precision and F1 score of the multiple operations after the agent office system provides the required content based on the second search instruction in the multiple operations;

[0055] A normalization processing unit, used for normalizing the accuracy, the recall, the precision and the F1 score;

[0056] A correction parameter acquisition unit, configured to determine a correction parameter set based on a result of the normalization processing, wherein the correction parameter set includes a weighted average of normalized values ​​of each of the multiple information items in the accuracy, the recall, the precision, and the F1 score;

[0057] A model confidence calculation unit is used to determine the current confidence of the preset instruction conversion model based on the historical model confidence and the correction parameter set.

[0058] In one embodiment of the present application, optionally, the model confidence calculation unit is used to:

[0059] The variance of the correction parameter set is obtained, and the difference between the variance and the historical model confidence is set as the current confidence of the preset instruction conversion model.

[0060] In a third aspect, an embodiment of the present application provides a computer device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are configured to execute the method described in the first aspect above.

[0061] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to execute the method described in the first aspect above.

[0062] The above technical solution, in view of the technical problem that the function retrieval method provided by the office system in the related technology is inefficient and inaccurate, only needs the user to provide a first retrieval instruction, and the first retrieval instruction can be converted into a second retrieval instruction readable by the intelligent office system in combination with the large language model, and the intelligent office system is driven to automatically read the second retrieval instruction and retrieve the required content required by the user. As a result, the user does not need to perform complex operations such as writing or searching for SQL query statements, and can directly complete the steps of searching for the required function in the intelligent office system through a simple retrieval instruction, which effectively improves the convenience of function retrieval in the office system and reduces the cost of manual operation. At the same time, although only one retrieval instruction is required from the user, the large language model is introduced in the technical solution to automatically convert the retrieval instruction into a form readable by the intelligent office system, avoiding the problem that the retrieval instruction provided by the user is inaccurate and affects the retrieval effect, and improving the accuracy of retrieval in the office system. In general, the technical solution can take into account the convenience and accuracy of internal retrieval in the office system, improve the retrieval efficiency, and simultaneously reduce the difficulty and cost of retrieval.

Brief Description of the Drawings

[0063] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0064] Figure 1 A flow chart showing a function retrieval method for an intelligent office system according to an embodiment of the present application is shown;

[0065] Figure 2 A flow chart showing a function retrieval method for an intelligent office system according to another embodiment of the present application is shown;

[0066] Figure 3 A block diagram of a computer device according to an embodiment of the present application is shown;

[0067] Figure 4 A block diagram of a computer device according to another embodiment of the present application is shown. [Specific implementation method]

[0068] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0069] Figure 1 A flow chart of a function retrieval method for an intelligent office system according to an embodiment of the present application is shown.

[0070] like Figure 1 As shown, a function retrieval method for an intelligent office system in one embodiment of the present application includes:

[0071] Step 102, obtaining a first search instruction for the intelligent office system.

[0072] An intelligent office system refers to an office system that has the functions of autonomous understanding, planning and decision-making, and executing complex tasks. In order to retrieve the functions they need in the intelligent office system, users can enter a first search instruction.

[0073] In a possible design, the way of inputting the first search instruction includes but is not limited to any user-implemented input method such as voice input, gesture input, text input, etc. It can be seen that the first search instruction is used to reflect the user's demand content for the intelligent office system.

[0074] In a possible design, the required content includes executing the operation action requested by the first search instruction and / or outputting the output content requested by the first search instruction. For example, the user can issue a voice instruction "settlement of this quarter's tax report" as an operation action, or directly input "this quarter's tax report" as the output content required by the user in the search position of the intelligent office system.

[0075] Step 104: convert the first search instruction into input information of a preset instruction conversion model.

[0076] The preset instruction conversion model is used to reflect the association between the demand content and the second search instruction for obtaining the demand content in the intelligent office system. The demand content refers to the content that the user needs to search for in the intelligent office system as indicated by the first search instruction, and the second search instruction refers to the actual instruction used by the intelligent office system to search for the function required by the user's first search instruction. In other words, the second search instruction is an instruction that the intelligent office system can understand.

[0077] The preset instruction conversion model is any large language model with natural language processing function, such as a pre-trained language model, a neural network language model, a deep learning language model or a hybrid language model. Optionally, the preset instruction conversion model is a baichuan-7b large model.

[0078] Specifically, since the first search instruction is the content directly input by the user into the intelligent office system, it is often irregular, that is, it may not match the instructions used by the intelligent office system when searching for its own functions. Therefore, the first search instruction can be converted into content consistent with the instructions used by the intelligent office system when searching for its own functions through a preset instruction conversion model.

[0079] Furthermore, before using the preset instruction conversion model, the first search instruction may be standardized in advance and converted into more standardized model input information to adapt to the function of the preset instruction conversion model.

[0080] In a possible design, the specific method of converting the first search instruction into input information in step 104 includes: extracting entities in the first search instruction as input information of the preset instruction conversion model.

[0081] The entity in the first search instruction refers to the subjects such as actions and operation objects mentioned in the first search instruction. For example, if the first search instruction is "export the corporate credit data of corporate user a in the form of a mixture of charts and graphs", the entities include "corporate user a", "corporate credit data" and "charts and graphs".

[0082] In this way, only the entity in the first search instruction is used as the input information of the preset instruction conversion model, and the redundant content in the first search instruction is deleted, which simplifies the input content of the preset instruction conversion model to a certain extent and reduces the calculation difficulty of the preset instruction conversion model. Furthermore, when pre-training the model, only the entity can be used as the input sample for training, which also simplifies the input sample and reduces the training difficulty of the model.

[0083] In another possible design, the specific method of converting the first search instruction into input information in step 104 includes: adding preset prefix information to the first search instruction to obtain input information of a preset instruction conversion model, wherein the preset prefix information is used to describe the processing requirements for converting the first search instruction into a second search instruction readable by the intelligent office system.

[0084] Compared with entity extraction, directly adding preset prefix information to the first search instruction is more convenient in the search process. For example, the preset prefix information can be set to "convert the first search instruction into a form readable by the intelligent office system", that is, this conversion instruction is added before the first search instruction. In this way, for the large language model, the first search instruction can be directly converted based on this conversion instruction.

[0085] On this basis, this technical solution can directly adopt the existing large language model. For a large number of retrieval instructions serving as input samples, the preset prefix information is added and then input into the large language model, thereby realizing the training and learning of the large language model for the function of converting readable instructions of the intelligent office system.

[0086] Step 106: Determine the second search instruction based on the input information and the preset instruction conversion model.

[0087] Step 108: input the second search instruction into the intelligent office system, so that the intelligent office system provides the required content based on the second search instruction.

[0088] At this point, the input information is input into the preset instruction conversion model, so that the preset instruction conversion model processes the input information and outputs a second search instruction readable by the intelligent office system. Then the second search instruction is input into the intelligent office system, and after the intelligent office system reads the obtained second search instruction, it automatically performs a search operation to obtain the required content required by the user.

[0089] The above technical solution only requires the user to provide a first search instruction, and the first search instruction can be converted into a second search instruction readable by the intelligent office system in combination with a large language model, and the intelligent office system is driven to automatically read the second search instruction and then retrieve the required content required by the user. As a result, the user does not need to perform complex operations such as writing or searching for SQL query statements, and can directly complete the steps of searching for the required functions in the intelligent office system through a simple search instruction, which effectively improves the convenience of function retrieval in the office system and reduces the cost of manual operation. At the same time, although the user only needs to provide one search instruction, the large language model is introduced in this technical solution to automatically convert the search instruction into a form readable by the intelligent office system, avoiding the problem that the search instruction provided by the user is inaccurate and affects the search effect, and improving the accuracy of the search in the office system. In general, this technical solution can take into account the convenience and accuracy of the internal search of the office system, improve the search efficiency, and simultaneously reduce the difficulty and cost of the search.

[0090] In a possible design, before step 106, it also includes: performing data structuring processing on the input information to obtain updated input information matching a preset structured format; then step 106 includes: determining the second retrieval instruction based on the updated input information and the preset instruction conversion model.

[0091] Specifically, after converting the first search instruction into input information that is more suitable for the preset instruction conversion model, further structural changes can be made to the input information, so that the structure of the updated input information is more regular, thereby reducing the processing difficulty of the preset instruction conversion model. Of course, when pre-training the model, the input samples after structured processing can also be used for training, thereby reducing the training difficulty of the model.

[0092] In one possible design, the process of training a preset instruction conversion model includes: training an initial instruction conversion model based on an initial sample data set; obtaining a dynamically updated real-time sample data set; in response to each update operation of the real-time sample data set, adjusting the top-level parameters of the initial instruction conversion model while freezing the bottom-level parameters of the initial instruction conversion model based on the updated real-time sample data set to obtain an iterative instruction conversion model; after the number of adjustments reaches a preset upper limit, setting the last iterative instruction conversion model as the preset instruction conversion model.

[0093] The model can be updated and trained in real time through the dynamically updated real-time sample data set. Each time the intelligent office system provides the required content based on the second search instruction, the data of this operation can be added to the dynamically updated real-time sample data set to achieve real-time update of samples. If the required content provided this time is correct, it is added to the positive sample set of the real-time sample data set, otherwise, it is added to the negative sample set.

[0094] Furthermore, in the first iteration for fine-tuning the model, the Lora technology can be used to freeze the bottom-level parameters of the initial instruction conversion model, and adjust the top-level parameters of the initial instruction conversion model at the same time. Among them, the bottom layer of the initial instruction conversion model has many common features, which are useful for different tasks. By freezing the bottom-level parameters, these common features can be migrated to the current task of retrieving instruction conversion in the office system, thereby providing a high-information input for the top layer of the initial instruction conversion model, which is conducive to improving the performance of the model in the current task, in other words, improving the matching degree between the model and the current task. At the same time, freezing some parameters can also prevent the model from overfitting and reduce the amount of calculation in the model training process, which helps to improve the model accuracy and reduce the difficulty of training. Thereafter, the process of each iteration is consistent with the above content and will not be repeated here.

[0095] In a possible design, after the intelligent office system provides the required content based on the second retrieval instruction in multiple operations, the accuracy, recall, precision and F1 scores of the multiple operations can be determined; the accuracy, recall, precision and F1 score can be normalized; based on the result of the normalization, a correction parameter set can be determined, wherein the correction parameter set includes the weighted average of the normalized values ​​of each item of information in the accuracy, recall, precision and F1 score; based on the historical model confidence and the correction parameter set, the current confidence of the preset instruction conversion model can be determined.

[0096] Among them, accuracy refers to the proportion of times the demand content retrieved in multiple searches is consistent with user needs, recall refers to the proportion of times the demand content retrieved in multiple searches is inconsistent with user needs, precision refers to the degree of match between the demand content retrieved in a single search and user needs, and F1 score refers to the harmonic mean of precision and recall, which is a comprehensive indicator compatible with precision and recall.

[0097] Specifically, after normalizing the accuracy, recall, precision and F1 score, the weights of each of the multiple information are obtained, and the weighted average of the multiple information is calculated. The weighted average reflects the reliability of the retrieval function of the intelligent office system under the comprehensive influence of the multiple information. It can be seen that the correction parameter set reflects the reliability of the retrieval function of the intelligent office system under the influence of various information and different combinations of various information.

[0098] Furthermore, determining the current confidence of the preset instruction conversion model based on the historical model confidence and the correction parameter set includes: obtaining the variance of the correction parameter set; and setting the difference between the variance and the historical model confidence as the current confidence of the preset instruction conversion model.

[0099] It should be noted that the current confidence of the preset instruction conversion model refers to the reliability of the preset instruction conversion model. The higher the current confidence, the lower the subsequent fine-tuning of the preset instruction conversion model. Conversely, the lower the current confidence, the higher the subsequent fine-tuning of the preset instruction conversion model.

[0100] Further, based on the current confidence of the preset instruction conversion model, the number of bottom-level parameters of the iterative instruction conversion model to be frozen and the number of top-level parameters of the iterative instruction conversion model to be adjusted during the iteration process can be determined. The higher the current confidence, the higher the number of bottom-level parameters of the iterative instruction conversion model to be frozen and / or the higher the number of top-level parameters of the iterative instruction conversion model to be adjusted.

[0101] Optionally, a plurality of sub-ranges may be divided within the distribution range of the current confidence, each sub-range corresponding to a set of bottom-level parameters that can be frozen and a set of top-level parameters that can be adjusted. After determining the current confidence of the preset instruction conversion model, the sub-range to which the current confidence belongs may be further determined, and then the bottom-level parameters that can be frozen and the top-level parameters that can be adjusted corresponding to the sub-range may be set as the bottom-level parameters that need to be frozen and the top-level parameters that need to be adjusted in the iterative process of fine-tuning the preset instruction conversion model.

[0102] Figure 2 A flow chart of a function retrieval method for an intelligent office system according to another embodiment of the present application is shown.

[0103] like Figure 2As shown, first, office data can be collected, and the collected data can be cleaned and normalized for preliminary training of the general data set. Furthermore, the model can be fine-tuned and evaluated based on the results of the preliminary training. During the model fine-tuning process, the underlying parameters of the model can be frozen and the upper parameters of the model can be fine-tuned. After fine-tuning, the fine-tuned model is evaluated on the validation set, and its evaluation indicators include accuracy, recall, precision, and F1 score. Next, the training parameters can be dynamically adjusted based on the evaluation results, and then optimized fine-tuning can be achieved based on the dynamically adjusted training parameters, and the step of evaluating the fine-tuned model on the validation set is returned to re-evaluate.

[0104] After multiple iterations, the hyperparameters are optimized through random search and Bayesian optimization, and a comprehensive performance evaluation of the model is performed after optimization. Of course, the process of hyperparameter optimization and comprehensive performance evaluation can also be iterated multiple times.

[0105] At this point, the model training for converting retrieved content for the office system has been completed.

[0106] On this basis, the user inputs the first search instruction to query, and the intelligent office system extracts key entities from the first search instruction, and uses the key entities as the input of the above model to identify the user's intention and manage the dialogue flow, and outputs a second search instruction that can be read by the intelligent office system. Finally, the intelligent office system executes the second search instruction, obtains the search results, and responds to the user with the search results. Among them, managing the dialogue flow includes multiple functions such as intention recognition, slot filling, response generation, and context tracking.

[0107] Specifically, in the technical solution of the present invention, the baichuan-7b large model is fine-tuned, and the Agent intelligent body technology is used to achieve efficient processing and understanding of internal office data of financial companies, and support material retrieval and multi-task parallel processing. In the core scenario of material retrieval, the vector database FAISS can be used for data storage and management to achieve efficient storage, retrieval and calculation of large-scale data. The vector database can perform fast search and nearest neighbor query on high-dimensional vectors, which is suitable for processing the ever-expanding text data. In the dimension of intelligent and parallel task processing, it can interact with users in natural language based on semantic understanding through interface dialogue, understand user query intentions, and disassemble the tasks by the Agent, and send user-processable requests in parallel.

[0108] Specifically, during the model training and optimization process, we can select the pre-trained model baichuan-7b suitable for processing Chinese text, and adjust it in combination with the specific needs of the financial field to ensure that the model has strong semantic understanding and generation capabilities and can adapt to the complex data of the financial industry. At the same time, we can use general data sets to conduct preliminary training on the pre-trained model.

[0109] In the process of model fine-tuning, in order to improve training efficiency and effectiveness, save training resources and reduce the complexity of the training process, the Lora technology can be used to freeze the underlying parameters of the pre-trained model and only adjust the top-level parameters. Specifically, the Lora technology can achieve this by adding low-rank adaptation layers, which are learned and adjusted according to domain-specific data during model training. This ensures that the model focuses on learning knowledge in specific fields while retaining general language understanding capabilities.

[0110] In addition, domain-specific datasets can be divided into training, validation, and test sets to ensure data balance and diversity, and fine-tuned using annotated data in the domain corpus to ensure that the model can learn terminology and semantics specific to financial industry companies.

[0111] During the training process, the above steps are iterated repeatedly to continuously optimize the model parameters so that the model's prediction results gradually approach the actual labels. According to the performance of the model on the validation set, the training parameters such as learning rate, batch size and number of training rounds can also be dynamically adjusted, and the learning rate scheduler can be used to adjust the learning rate during the training process to ensure that the model can converge better.

[0112] At the same time, in order to prevent the model from overfitting, regularization Dropout and L2 regularization techniques can be added during the training process. By regularizing, a part of neurons can be randomly discarded to prevent the model from over-relying on certain specific neurons and improve the generalization ability of the model.

[0113] For training data construction and large model fine-tuning, first, collect the full amount of data of various types of company office domain materials, such as unit information, onboarding information, project management and contracts. Then, clean the collected data and normalize the text to remove noise, duplication and incomplete data to ensure data quality and consistency. Then, use natural language processing technology to annotate the data, including named entity recognition, text classification and sentiment analysis, to generate high-quality training and test sets. At the same time, unstructured data (such as text) can also be converted into structured formats (such as JSON, CSV) for effective data processing by large models. Then, use the domain corpus to fine-tune the pre-trained model, freeze some layers of the pre-trained model, and adjust the parameters of the top layer so that the model can understand the specific knowledge and semantics within the company. Finally, use the validation set and test set to evaluate the fine-tuned model. Among them, the evaluation indicators include accuracy, recall, precision and F1 score. Through these evaluation indicators, we can fully understand the performance of the model in different dimensions. Based on the evaluation results, random search and Bayesian optimization are used to optimize the model's overall hyperparameters. In the predefined parameter space, a portion of hyperparameter combinations are randomly selected for evaluation. At the same time, a proxy model can be constructed to predict the performance of different hyperparameter combinations and select the optimal hyperparameter combination.

[0114] For the modular design of the office system architecture, first of all, the system architecture is divided into the data layer, model layer and application layer. The data layer is responsible for data storage and management, and supports vector databases and storage of multiple data formats. The model layer performs semantic understanding and information generation based on the fine-tuned large model, realizes functions such as text classification, information extraction, and question-answering system, and supports the processing and analysis of complex data. Through the semantic processing capabilities of the model layer, the system's understanding and response capabilities to user queries can be improved. The model layer can also provide model services to support real-time semantic parsing and task processing. The model service interacts with the application layer through the API interface to process user query requests and return results. The application layer provides user interfaces and interactive functions, supports multiple query methods and task management, realizes the parsing and processing of user input, and can also call the services of the model layer and data layer to generate query results and display them to users. Through the collaborative work of the front-end and back-end, it ensures that users can quickly obtain the required information.

[0115] Regarding the application of agent technology, in detail, intelligent agent is an important part of the intelligent office system of the present invention, responsible for natural language interaction with users, understanding user intentions, and executing corresponding tasks. Based on the fine-tuned large model, the intelligent agent has the ability of semantic understanding and intention recognition, can parse the user's natural language query, and convert it into a request that the system can process. In other words, natural language processing technology and dialogue management framework are used to realize the dialogue management and interaction control of intelligent agents.

[0116] Among them, the intelligent agent needs to have strong semantic parsing capabilities. Optionally, NER technology can be used to extract key entities in the text, such as names of people, places, institutions, dates, etc., and then a large amount of entity data can be trained to enable the model to accurately identify the user's intentions, such as querying information, executing tasks, etc. Based on the key entities, the intelligent office system can better understand the user's query content. For example, when querying "Please help me find contracts related to finance", the intelligent office system can identify that finance is the definition domain of the contract, and querying the contract list is the user's true intention.

[0117] In addition, dialogue management is an important part of the user experience of the intelligent office system. Specifically, the dialogue management framework Dialogflow can be used to manage the interaction process with users, including intent recognition, slot filling, and response generation. The dialogue management framework ensures that the intelligent office system can handle complex multi-round dialogues by defining the dialogue flow and intent recognition logic and tracking the context and status of the user session. The intelligent office system sends the context of the dialogue together with the parsed executable intent to the big model to execute the user's required operation.

[0118] The embodiment of the present application provides an intelligent operation device for an intelligent office system, including:

[0119] A first search instruction acquisition unit is used to acquire a first search instruction for the intelligent office system, wherein the first search instruction is used to reflect the user's demand content for the intelligent office system, and the demand content includes executing the operation action requested by the first search instruction and / or outputting the output content requested by the first search instruction;

[0120] An input information determination unit, configured to convert the first search instruction into input information of a preset instruction conversion model, wherein the preset instruction conversion model is used to reflect the association relationship between the demand content and a second search instruction for obtaining the demand content in the agent office system;

[0121] A second search instruction acquisition unit, configured to determine the second search instruction based on the input information and the preset instruction conversion model;

[0122] The required content output unit is used to input the second search instruction into the intelligent agent office system, so that the intelligent agent office system provides the required content based on the second search instruction.

[0123] In one embodiment of the present application, optionally, the input information determining unit includes:

[0124] The first execution unit is used to extract entities in the first search instruction as input information of the preset instruction conversion model.

[0125] In one embodiment of the present application, optionally, the input information determining unit includes:

[0126] The second execution unit is used to add preset prefix information to the first search instruction to obtain input information of the preset instruction conversion model, wherein the preset prefix information is used to describe the processing requirements for converting the first search instruction into a second search instruction readable by the intelligent office system.

[0127] In one embodiment of the present application, optionally, the preset instruction conversion model is a pre-trained language model, a neural network language model, a deep learning language model or a hybrid language model.

[0128] In one embodiment of the present application, optionally, the device further includes:

[0129] An input information structured processing unit, used for performing data structured processing on the input information before determining the second search instruction, to obtain updated input information matching a preset structured format;

[0130] The second search instruction acquisition unit is used for:

[0131] The second search instruction is determined based on the updated input information and the preset instruction conversion model.

[0132] In one embodiment of the present application, optionally, the device further includes:

[0133] An initial model training unit, used for training an initial instruction conversion model based on an initial sample data set before determining the second retrieval instruction;

[0134] A real-time sample data set acquisition unit, used to acquire a dynamically updated real-time sample data set;

[0135] a model parameter optimization unit, configured to adjust the top-level parameters of the initial instruction conversion model in response to each update operation of the real-time sample data set and based on the updated real-time sample data set, while freezing the bottom-level parameters of the initial instruction conversion model, to obtain an iterative instruction conversion model;

[0136] The model determination unit is used to set the last iteration instruction conversion model as the preset instruction conversion model after the adjustment times reach the preset upper limit.

[0137] In one embodiment of the present application, optionally, the device further includes:

[0138] A feedback information determination unit, configured to determine the accuracy, recall, precision and F1 score of the multiple operations after the agent office system provides the required content based on the second search instruction in the multiple operations;

[0139] A normalization processing unit, used for normalizing the accuracy, the recall, the precision and the F1 score;

[0140] A correction parameter acquisition unit, configured to determine a correction parameter set based on a result of the normalization processing, wherein the correction parameter set includes a weighted average of normalized values ​​of each of the multiple information items in the accuracy, the recall, the precision, and the F1 score;

[0141] A model confidence calculation unit is used to determine the current confidence of the preset instruction conversion model based on the historical model confidence and the correction parameter set.

[0142] In one embodiment of the present application, optionally, the model confidence calculation unit is used to:

[0143] The variance of the correction parameter set is obtained, and the difference between the variance and the historical model confidence is set as the current confidence of the preset instruction conversion model.

[0144] The device uses any one of the solutions described in the above embodiments, and therefore has all the above technical effects, which will not be repeated here.

[0145] In addition, in one embodiment, the present application provides a computer device, which may be a server, and its internal structure diagram may be as follows: Figure 3As shown. The computer device includes a processor, a memory, a network interface and a database connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile and / or volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external client via a network connection. When the computer program is executed by the processor, the method described in any of the above embodiments can be implemented.

[0146] In one embodiment, the present application further provides a computer device, which may be a client, and its internal structure diagram may be as follows: Figure 4 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external server via a network connection. When the computer program is executed by the processor, the method described in any of the above embodiments can be implemented.

[0147] Any of the above-mentioned computer devices in the embodiments of the present application may exist in various forms, including but not limited to:

[0148] (1) Mobile communication devices: These devices are characterized by their mobile communication functions and their main purpose is to provide voice and data communications. These terminals include: smart phones (such as iPhone), multimedia phones, functional phones, and low-end phones.

[0149] (2) Ultra-mobile personal computer devices: These devices fall into the category of personal computers, have computing and processing capabilities, and generally also have mobile Internet access features. These terminals include: PDA, MID and UMPC devices, such as iPad.

[0150] (3) Portable entertainment devices: These devices can display and play multimedia content. They include audio and video players (such as iPods), handheld game consoles, e-books, as well as smart toys, wearable devices, and portable car navigation devices.

[0151] (4) Server: A device that provides computing services. The server consists of a processor, hard disk, memory, system bus, etc. The server is similar to a general computer architecture, but because it needs to provide highly reliable services, it has higher requirements in terms of processing power, stability, reliability, security, scalability, and manageability.

[0152] (5) Other electronic devices with data interaction functions.

[0153] In addition, an embodiment of the present application provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to perform the following steps:

[0154] Acquire a first search instruction for the intelligent office system, wherein the first search instruction is used to reflect the user's demand content for the intelligent office system, and the demand content includes executing the operation action requested by the first search instruction and / or outputting the output content requested by the first search instruction;

[0155] Converting the first search instruction into input information of a preset instruction conversion model, wherein the preset instruction conversion model is used to reflect the association relationship between the demand content and a second search instruction for obtaining the demand content in the agent office system;

[0156] Determining the second search instruction based on the input information and the preset instruction conversion model;

[0157] The second search instruction is input into the intelligent agent office system, so that the intelligent agent office system provides the required content based on the second search instruction.

[0158] It should be noted that the above functions or steps that can be implemented by the computer-readable storage medium or computer device can refer to the relevant description in the aforementioned method embodiment. In order to avoid repetition, they will not be described one by one here.

[0159] The technical solution of the present application is described in detail above in combination with the accompanying drawings. Through the technical solution of the present application, the user only needs to provide a first search instruction, and the first search instruction can be converted into a second search instruction readable by the intelligent office system in combination with the large language model, and the intelligent office system is driven to automatically read the second search instruction and retrieve the required content required by the user. As a result, the user does not need to perform complex operations such as writing or searching for SQL query statements, and the steps of searching for the required functions in the intelligent office system can be directly completed through a simple search instruction, which effectively improves the convenience of function retrieval in the office system and reduces the cost of manual operation. At the same time, although the user only needs to provide one search instruction, the large language model is introduced in the technical solution to automatically convert the search instruction into a form readable by the intelligent office system, avoiding the problem that the search instruction provided by the user is inaccurate and affects the search effect, and improving the accuracy of the search in the office system. In general, the technical solution can take into account the convenience and accuracy of the internal search of the office system, improve the search efficiency, and simultaneously reduce the difficulty and cost of the search.

[0160] It should be understood that the term "and / or" used in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0161] It should be understood that, although the terms first, second, etc. may be used to describe the retrieval instruction in the embodiments of the present application, these retrieval instructions should not be limited to these terms. These terms are only used to distinguish the retrieval instructions from each other. For example, without departing from the scope of the embodiments of the present application, the first retrieval instruction may also be referred to as the second retrieval instruction, and similarly, the second retrieval instruction may also be referred to as the first retrieval instruction.

[0162] The word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to determining" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)", depending on the context.

[0163] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the present application. The singular forms "a", "said" and "the" used in the embodiments of the present application and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings.

[0164] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0165] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0166] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0167] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A function retrieval method for an intelligent office system, characterized in that: include: Acquire a first search instruction for the intelligent office system, wherein the first search instruction is used to reflect the user's demand content for the intelligent office system, and the demand content includes executing the operation action requested by the first search instruction and / or outputting the output content requested by the first search instruction; Converting the first search instruction into input information of a preset instruction conversion model, wherein the preset instruction conversion model is used to reflect the association relationship between the demand content and a second search instruction for obtaining the demand content in the agent office system; Determining the second search instruction based on the input information and the preset instruction conversion model; The second search instruction is input into the intelligent agent office system, so that the intelligent agent office system provides the required content based on the second search instruction.

2. The method according to claim 1, characterized in that The step of converting the first search instruction into input information of a preset instruction conversion model includes: Entities in the first search instruction are extracted as input information of the preset instruction conversion model.

3. The method according to claim 1, characterized in that The step of converting the first search instruction into input information of a preset instruction conversion model includes: Preset prefix information is added to the first search instruction to obtain input information of a preset instruction conversion model, wherein the preset prefix information is used to describe the processing requirements for converting the first search instruction into a second search instruction readable by the intelligent office system.

4. The method according to any one of claims 1 to 3, characterized in that The preset instruction conversion model is a pre-trained language model, a neural network language model, a deep learning language model or a hybrid language model.

5. The method according to any one of claims 1 to 3, characterized in that Before determining the second search instruction based on the input information and the preset instruction conversion model, the method further includes: Performing data structuring processing on the input information to obtain updated input information matching a preset structured format; The determining the second search instruction based on the input information and the preset instruction conversion model includes: The second search instruction is determined based on the updated input information and the preset instruction conversion model.

6. The method according to claim 1, characterized in that Before determining the second search instruction based on the input information and the preset instruction conversion model, the method further includes: Based on the initial sample data set, train the initial instruction conversion model; Get a dynamically updated real-time sample dataset; In response to each update operation of the real-time sample data set, based on the updated real-time sample data set, adjusting the top-level parameters of the initial instruction conversion model while freezing the bottom-level parameters of the initial instruction conversion model to obtain an iterative instruction conversion model; After the number of adjustments reaches the preset upper limit, the last iteration instruction conversion model is set as the preset instruction conversion model.

7. The method according to claim 6, characterized in that Also includes: After the agent office system provides the required content based on the second search instruction in multiple operations, determining the accuracy, recall, precision and F1 score of the multiple operations; Normalizing the accuracy, the recall, the precision, and the F1 score; Based on the result of the normalization processing, determining a set of correction parameters, wherein the set of correction parameters includes a weighted average of normalized values ​​of each of the multiple information in the accuracy, the recall, the precision and the F1 score; Based on the historical model confidence and the set of correction parameters, a current confidence of the preset instruction conversion model is determined.

8. The method according to claim 7, characterized in that The determining the current confidence of the preset instruction conversion model based on the historical model confidence and the correction parameter set includes: Obtaining the variance of the correction parameter set; The difference between the variance and the historical model confidence is set as the current confidence of the preset instruction conversion model.

9. A computer device, characterized in that: include: at least one processor; and, a memory communicatively coupled to the at least one processor; The memory stores instructions executable by the at least one processor, wherein the instructions are configured to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: Computer executable instructions are stored, and the computer executable instructions are used to execute the method according to any one of claims 1 to 7.